🧪 Test?View on arXiv
DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs
Author1, Author2, Author3, Author4, Author5
inferencehardware repurposingenergy efficiency
2608.14614
Builder Relevance
1h ago80%
Abstract
This paper investigates the economic viability and environmental sustainability of repurposing retired GPUs to form a DumpsterCluster for modern LLM inference.
Reality Card
Core Claim
The authors built a 128-GPU DumpsterCluster that achieves competitive LLaMA-70B throughput at a significantly lower cost compared to traditional systems.
Method / Result
Achieved competitive throughput for LLaMA-70B using a 128-GPU setup costing $22K.
Limitations
Older GPUs have higher energy consumption, making the total cost of ownership favorable only in regions with inexpensive electricity.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.
No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.